复杂三维点云场景中的目标识别方法研究

发布时间：2018-03-18 11:32

本文选题：三维点云　切入点：局部特征　出处：《吉林大学》2017年硕士论文　论文类型：学位论文

【摘要】：三维点云中的目标识别一直是计算机图像和模式识别领域中的一个热门话题,然而三维点云中存在着如噪声污染、目标遮挡和点云密度变化等滋扰现象,这些滋扰增加了目标识别的难度。基于局部特征的目标识别可以很好地解决复杂场景中的目标遮挡问题,但是目前对处理噪声尤其是点云密度变化的相关研究还比较少。为了解决复杂三维点云场景中目标识别出现的问题,本文主要完成了以下工作:首先概述了基于局部特征的三维目标识别方法,详细介绍了每一个步骤的具体方法。并对影响目标识别效率的两个关键步骤——特征点和特征描述提取进行了大量研究。在特征点提取方面,主要研究了目前识别效果较好的三维固有形状签名(Intrinsic Shape Signatures 3D,ISS3D)特征点方法,通过采取带距离和点云密度权重的散布矩阵(Scatter Matrix)处理ISS3D对噪声和点云密度变化特别敏感的问题。同时利用张量投票的方法去除离群点,最后用表面曲率作为显著度增加特征点数目。通过在Retrieval模型库和Bo D1模型库中进行的特征点性能比较实验,证明了改进后的ISS3D特征点方法对噪声和点云密度变化的具有较高的鲁棒性。在特征描述阶段,主要研究了目前点云特征中比较常用的方向直方图签名特征(Signature of Histograms of Orien Tations,SHOT)。SHOT特征继承了基于签名特征的高描述性,同时具备了基于直方图特征的高鲁棒性,在三维点云识别中应用广泛。SHOT特征对噪声具有很高的鲁棒性,但极易受点云密度变化影响。针对这个问题,本文在建立局部坐标系(Local Reference Frame,LRF)时提出了一种新的几何重心并添加了点云密度权重。在计算SHOT直方图特征时,提出了一种新的计算法向的方法,能够较好的处理噪声和点云密度变化。实验发现改进后的SHOT特征提高了对噪声和点云密度变化描述性与鲁棒性。在复杂场景中的三维目标识别实验中,本文首先对Bo D1模型库中进行了降采样,生成了实验所需的同时存在噪声和点云密度变化问题的模型库。在进行目标识别实验时,本文使用了最近邻比次近邻的方法(NNDR)计算初步的特征匹配,然后利用广义霍夫变换(Generalized Hough Transform,GHT)生成位置假设,最后使用ICP算法对所有假设进行验证,从中选择出目标复杂点云场景中目标可能出现的位置。实验表明,对于复杂的三维点云场景中的目标识别,本文的方法对于噪声和点云密度变化问题具有很好的识别效果。
[Abstract]:Target recognition in 3D point cloud has been a hot topic in the field of computer image and pattern recognition. However, there are nuisance phenomena such as noise pollution, target occlusion and change of point cloud density in 3D dot cloud. These disturbances increase the difficulty of target recognition. Target recognition based on local features can solve the problem of target occlusion in complex scenes. In order to solve the problem of target recognition in complex 3D point cloud scene, however, there are few researches on how to deal with the noise, especially the change of point cloud density. The main work of this paper is as follows: firstly, the method of 3D target recognition based on local features is summarized. The specific methods of each step are introduced in detail. Two key steps that affect the efficiency of target recognition, feature point extraction and feature description extraction, are studied. In this paper, we mainly study the feature point method of intrinsic Shape Signatures 3DX ISS3D, which has good recognition effect at present. A scatter matrix with distance and weight of point cloud density is adopted to deal with the problem that ISS3D is particularly sensitive to noise and point cloud density changes. At the same time, the method of Zhang Liang voting is used to remove outliers. Finally, the surface curvature is used to increase the number of feature points. The performance of feature points in Retrieval model base and Bo D1 model base is compared with each other. It is proved that the improved ISS3D feature point method is robust to the change of noise and point cloud density. In this paper, the signature of Histograms of Orien histogram signature signature of Orien histogram signature is studied, which inherits the high description based on signature feature and has high robustness based on histogram feature. The SHOT feature, which is widely used in 3D point cloud recognition, is highly robust to noise, but easily affected by the change of point cloud density. In this paper, we propose a new geometric center of gravity and add the weight of point cloud density in the local coordinate system of Local Reference frame. A new method for calculating the normal direction of SHOT histogram is presented. The experimental results show that the improved SHOT features improve the description and robustness of noise and point cloud density changes. In this paper, we first sampled the Bo D1 model base and generated the model base, which has the problem of noise and point cloud density change simultaneously in the experiment. In this paper, the nearest neighbor ratio (NNDR) method is used to calculate the initial feature matching, and then the generalized Hough transform is used to generate the location hypothesis. Finally, the ICP algorithm is used to verify all the assumptions. The experimental results show that the method proposed in this paper has a good effect on the noise and the change of point cloud density in the complex 3D point cloud scene.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】